search

theAIGuysCode / yolov3_deepsort

Object tracking implemented with YOLOv3, Deep Sort and Tensorflow.
GNU General Public License v3.0
337 stars 183 forks source link

Using yolov4 instead of v3 #9

Closed utkutpcgl closed 4 years ago

utkutpcgl commented 4 years ago

Recently the new version came out. Is there an easy way to integrate it with the code?

vihari1729 commented 4 years ago

@utkutpcgl Did you find any way integrate yolov4model(which has been trained on dataset) for tracking?

utkutpcgl commented 4 years ago

@vihari1729. I modified an existing github repository of yolov4 [https://github.com/hunglc007/tensorflow-yolov4-tflite] , implemented in tf. After integrating and modifying the model file it worked with @theAIGuysCode yolov3deepsort.

vihari1729 commented 4 years ago

@utkutpcgl Thanks, It is working now

MuhammadAsadJaved commented 3 years ago

@utkutpcgl Hi, How about the speed of YOLOV4+DeepSort? + Have you tried Yolov4-Tiny + DeepSort as well?

utkutpcgl commented 3 years ago

Hello @MuhammadAsadJaved. I did not try it with tiny. I tried it with v3 and v3 tiny. V4 was ~25% more successful than v3 in terms of flickering etc. However, v3-tiny could not match my expectations. I trained it on 1070ti with 2000 epoch. The batch size was 32 or 64, I do not remember exactly. I had very few sample images around 100 or so. I expected tiny to learn faster and generalize better, however, it ended up as v4 being the best. I cannot really comment on v4 tiny. Hopefully this helps you.

By the way you are welcome @vihari1729 .

MuhammadAsadJaved commented 3 years ago

@utkutpcgl How about the speed of V4+DeepSort?

utkutpcgl commented 3 years ago

Deep sort does not affect the fps much. The delay is mostly due to yolov4 on its own. I had around 9-10 fps on v3+deepsort and around 6-7 fps on v4+deepsort with ti 1070. The resolution was 416x416 if I remember correctly.